Combining Yolov3-tiny Model with Dropblock for Tiny-face Detection

Abstract

Face detection is one of the most basic tasks in many various face applications, which is gradually becoming the most acceptable biometric recognition method. However, tinyface detection is a complex and challenging pattern detection problem that encounters many difficulties in the application process. This paper uses the popular deep learning detection algorithm to complete tiny-face detection and combine the Yolov3-tiny model with Dropblock strategy. Dropblock is a regularization method used in the convolutional layer, in which the activation units are spatially interrelated. Neurons in the contiguous region of the feature map would be dropped by dropblock, which forces the neural network to learn novel features. Numerous experiments showed that Yolov3-tiny works better by combining with dropblock in tiny-face detection. We evaluate the performance of our proposed method on the two public datasets and increase 7.74% accuracy, compared with yolov3-tiny based model in the tinyface detection.